Web Intelligence Specification (WIS)

Metadata

Tag	Value
Proposal	2024-04-13-WIS
Authors	Vishal Pratap Singh, Harman Marshal Singh
Review Manager	TBD
Status	Draft
Implementations	NULL
Issues	NULL
Previous Revisions	NULL

Introduction

The Web Intelligence Specification (WIS) provides a set of guidelines for structuring web applications to improve interaction with AI agents.

Motivation

Creating LLM agents that browse the web is difficult.

Due to the lack of any structure or information, LLM agents need to be configured by the creators for every web application separately.

However, envision a paradigm shift: if the web applications themselves can help LLM agents browse them effectively. It would be a game-changer.

The WIS approach proposes a universal standard to effectively decouple the visiting AI Agent's understanding capabilities from the application's logic and user interface.

By doing so, it allows developers to build AI-friendly web applications without tailoring them to a specific AI’s capabilities.

Proposed solution

We propose a very simple guideline for web applications to follow.

Developers could implement these 2 workflows into their web application.

Workflow 1

In workflow 1 (Actions Workflow) the visiting agent has to run a loop to perform any operation.

It would be visiting /actions before performing any other operation.

This endpoint will give instructions to the visiting agent about how to should an agent start to navigate this web app.

Meanwhile, For the web applications following WIS workflow 1, it is mandated to implement a /actions route.

This route would return a JSON object which would contain:

• List of endpoints (Just L1 hierarchy)
• Description in natural language of their functionalities
• List of input parameters
• Return value details

{
  "status",
  "actions":[
    {
      endpoint,
      description,
      parameters:[],
      return_data_details
    }
  ],
}

//actions: 1st level heirarchy endpoints

Each of the endpoint which are mentioned in the /actions endpoint are mandated to have their return data in the following format.

They too would contain a list of endpoints instructing further actions that the agent could perform after it has the return data in this response.

It's a tree-like structure, similar to how users navigate through UI interfaces.

{
  "status",
  "data",
  "actions":[
    {
      endpoint,
      description,
      parameters:[],
      return_data_details
    }
  ],
}

// ${action_endpoint_x}: (n+1)th level heirarchy endpoints

Workflow 2

Workflow 2 (Ask Workflow) is recommended only for more complex web applications with the resources to create their own LLM application.

This is a more futuristic workflow, where one agent talks to the other.

Any workflow that implements this workflow is mandated to create a /ask route.

An external user agent could query /ask in natural language and get things done directly without running any loops.

This obviously means that the loop is run by the web application's own agent internally and all the processing has been shifted to the application side instead of users side.

This would take more effort to create, hence not recommended for smaller applications.

The /actions endpoint serves as a discovery mechanism for AI to ascertain what the service can do and then query endpoints in a loop, while the /ask endpoint provides a way for AI to engage with the application using natural language.

Comparison With Current Workarounds

Current workarounds are either trying to create an agent that could interact with frontend interfaces that are primarily built for humans.

Or are trying to integrate one web application at a time, which restricts small developers to build general-purpose agents and applications with ease.

Example

Placing orders from a nearby restaurant with good reviews.

In this example, we will demonstrate the actions workflow.

Backwards compatibility and Benefits

• Gradual Adoption: WIS allows for incremental implementation alongside existing features.

• Modular Approach: Acts as an independent module, leaving the core application unchanged.

• Utilization of Current APIs: Enhances current APIs with AI capabilities without fundamental changes.

• No Downtime Required: Integrates with current systems without service interruptions.

• Independent Updates: Permits isolated updates to the AI layer without affecting the main application.

• Optional Use: Provides AI functionalities as an optional layer for users who wish to utilize it.

• Phased Roll-Out: Enables controlled, step-wise integration to assess impact and performance.

---

P.S. - While this could have been just an extension of accessibility standards and it may merge into them. It is important to take into consideration that in the future there will be web applications that are built solely for AI agents. Such websites probably would have no HTML code and UI at all. So the standard has to be for the server side only unlike other common accessibility guidelines that are written mostly for the UI.